A geomagnetic modeling method based on fully connected deep neural network

The geomagnetic modeling method using a fully connected deep neural network with residual networks and genetic algorithm optimization addresses gradient issues, enhancing accuracy and spatial resolution in geomagnetic field modeling.

CN118690665BActive Publication Date: 2024-12-06NINGBO INST OF NORTHWESTERN POLYTECHNICAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411170774.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-12-06
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

Conventional spherical harmonic analysis for geomagnetic field modeling is limited by spatial resolution and the gradient vanishing or exploding issues in fully connected deep neural networks hinder accurate geomagnetic field modeling.

Method used

A geomagnetic modeling method using a fully connected deep neural network with residual networks and batch normalization, combined with a genetic algorithm to optimize hyperparameters, addresses the gradient issues and improves accuracy.

Benefits of technology

The method enhances the accuracy and convergence of geomagnetic field modeling by mitigating gradient vanishing or exploding problems, resulting in improved spatial resolution and model optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118690665B_ABST
    Figure CN118690665B_ABST
Patent Text Reader

Abstract

The present invention relates to a geomagnetic modeling method based on a fully connected deep neural network. Aiming at the problems that the existing geomagnetic model has low accuracy and long updating time of spherical harmonic parameters, the present invention obtains a geomagnetic field deep neural network model through joint optimization of a fully connected deep neural network, a residual network and a genetic algorithm, thereby solving the problems of low accuracy of traditional geomagnetic models and gradient explosion caused by increasing the number of layers of traditional fully connected networks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of satellite magnetic survey data modeling, and in particular to a geomagnetic modeling method based on a fully connected deep neural network. Background Art

[0002] The geomagnetic field is an important geophysical feature of the Earth. It carries rich information about the internal and external space of the Earth and is closely related to human production activities. It is widely used in many fields such as geophysical exploration, remote sensing positioning, planetary science research and social applications. Satellite magnetometry can quickly obtain three-component magnetic data with uniform global coverage in a short period of time. One of its main tasks is to use this to build a high-precision geomagnetic field model.

[0003] The conventional method of magnetic field modeling is the spherical harmonic analysis method proposed by Gauss, which uses the sum of a finite number of spherical harmonic functions (truncated spherical harmonic series) with increasing orders (increasing spatial resolution) to approximate the global geomagnetic field scalar potential and make the best fit to the global magnetic survey data. The spherical harmonic model is limited by the truncation level of the spherical harmonic series, and the spatial resolution is generally not high. When the truncation order increases, the number of spherical harmonic coefficients to be calculated increases sharply, and the coefficients of high-order terms are much smaller than those of low-order terms (the difference is as large as 6 orders of magnitude), and their statistical significance is questioned. Therefore, the traditional spherical harmonic series modeling is far from meeting the actual needs of prospecting and navigation.

[0004] There is a complex nonlinear relationship between the geomagnetic field and factors such as geographic location, elevation and time. The end-to-end learning ability of the fully connected deep network model allows the deep network to directly generate the final result from the raw data, which is very suitable for dealing with such nonlinear problems. However, in order to make the fully connected network model show higher accuracy in the learning task of geomagnetic field multi-source fusion data, increasing the depth of the network model is one of the effective strategies. This method is effective within a certain range, but simply increasing the depth of the network may cause a series of problems, especially gradient disappearance or gradient explosion. The gradient may gradually decrease to the point where it is almost impossible to update the weights, thereby hindering the further learning of the model. Even if the gradient disappearance or explosion problem is not so significant, it may cause the model to be difficult to converge, affecting the optimization of the final result. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to improve the gradient vanishing or exploding problem of a fully connected deep network model. In order to overcome the defects of the above-mentioned prior art, the present invention provides a geomagnetic modeling method based on a fully connected deep neural network.

[0006] The present invention provides a geomagnetic modeling method based on a fully connected deep neural network, comprising:

[0007] Step 1. Obtain satellite magnetic survey data, and divide the satellite magnetic survey data into training set and test set after preprocessing;

[0008] Step 2. Establish a geomagnetic field deep neural network model based on a residual network and a fully connected deep neural network; the geomagnetic field deep neural network model includes an input layer, four hidden layers and an output layer connected in sequence, the four hidden layers are connected in sequence, and the connection nodes between the input layer and the first hidden layer, between the first hidden layer and the second hidden layer, between the second hidden layer and the third hidden layer, between the third hidden layer and the fourth hidden layer, and between the fourth hidden layer and the output layer have weights, and a residual function and batch normalization processing are added to each of the weights; and the hyperparameters of the geomagnetic field deep neural network model are designed;

[0009] Step 3. Optimize the hyperparameters of the geomagnetic field deep neural network model based on mutation, crossover and selection of genetic algorithm; substitute the optimized hyperparameters into the geomagnetic field deep neural network model;

[0010] Step 4. Input the training set into the geomagnetic field deep neural network model in step 3 for training;

[0011] Step 5. Input the test set into the trained geomagnetic field deep neural network model for verification.

[0012] Compared with the prior art, the present application has the following advantages: In order to address the problems of low accuracy of existing geomagnetic models and long parameter updating time, the present invention obtains a geomagnetic field deep neural network model through joint optimization of a fully connected deep neural network, a residual network, and a genetic algorithm, thereby solving the problems of low accuracy of traditional geomagnetic models and gradient explosion caused by the deepening of the number of layers of traditional fully connected networks.

[0013] In a possible implementation, the characteristic elements of the satellite magnetic data in step 1 include the time of reading the magnetic data, the longitude and latitude and the spherical center distance of the magnetic data, and the three components of the magnetic field in the NEC coordinate system; the preprocessing of the satellite magnetic data in step 1 is specifically to filter the satellite magnetic data according to the interference label of the satellite magnetic data, the local time condition, and the geomagnetic index condition; and the filtered satellite magnetic data is normalized.

[0014] In a possible implementation, the input layer, hidden layer and output layer of the geomagnetic field deep neural network model in step 2 respectively include a plurality of neurons, and activation functions are respectively provided in the input layer, hidden layer and output layer, and each neuron in the input layer respectively inputs a feature element After adding the residual function and batch normalization processing of each weight and nonlinear activation function, the output of each neuron in the output layer is .

[0015] In a possible implementation manner, the adding of the residual function and the batch normalization processing to each weight in step 2 specifically include:

[0016] The output of the input layer and each hidden layer is , when processing input feature elements And each out Add the residual function , expressed as ; Construct the identity mapping of the geomagnetic field deep neural network model;

[0017] Each Normalize to an array with a mean of 0 and a variance of 1. The specific method is: , , ,in is the dimension of the output, is the mean, is the standard deviation, It is 10 orders of magnitude smaller to prevent the occurrence of 0 which causes the division to fail. is the output of the i-th layer, is the normalized output.

[0018] In one possible implementation, the hyperparameters for designing the geomagnetic field deep neural network model in step 2 include the number of samples, learning rate, reduction factor, training cycle, and patience.

[0019] In a possible implementation manner, the hyperparameters of the geomagnetic field deep neural network model optimized by mutation, crossover and selection based on the genetic algorithm in step 3 specifically include:

[0020] Step 301. Set the parameter ranges of the number of samples, learning rate, reduction factor, training cycle and patience, randomly select values ​​of the five hyperparameters to form individuals, and randomly select multiple individuals to form the initial population;

[0021] Step 302. Each parameter value is regarded as a gene and represented in the form of binary coding; the parameter values ​​of the five hyperparameters, namely, the number of samples, the learning rate, the reduction factor, the training cycle and the patience, are combined to form a chromosome, and each chromosome represents an individual;

[0022] Step 303. Use the fitness function to calculate the fitness value of the individual. The fitness function is expressed by the mean square error between the predicted value and the true value of the fully connected network. The smaller the mean square error, the closer it is to the true value.

[0023] Step 304: Sort the fitness values ​​in the initial population, and screen out individuals whose fitness values ​​are less than a preset threshold, and use a genetic algorithm to perform crossover, mutation, and linear recombination to obtain the next generation population;

[0024] Step 305. For the next generation population, adopt steps 303 and 304, loop to obtain the T-th generation population, select the value with the largest fitness from the T-th generation population as the optimal parameter of each hyperparameter; and substitute it into the geomagnetic field deep neural network model.

[0025] In a possible implementation manner, in step 4, inputting the training set into the geomagnetic field deep neural network model in step 3 for training specifically includes:

[0026] The step 4 specifically includes: inputting the training set into the geomagnetic field deep neural network model in step 3 for training, specifically includes: inputting the training set of step 1 into the geomagnetic field deep neural network model in step 305 for training, calculating the loss and updating; wherein the loss adopts the mean square error between the predicted value and the true value of the previous propagation; the updating includes a gradient descent process and a back propagation algorithm, adopting the Adam algorithm as the gradient descent algorithm, and adopting the back propagation algorithm to update the weights of the connection nodes between the input layer and the hidden layer, between the hidden layer and the hidden layer, and between the hidden layer and the output layer in the geomagnetic field deep neural network model.

[0027] In a possible implementation, the step 5 of inputting the test set into the trained geomagnetic field deep neural network model for verification specifically includes: inputting the test set into the trained geomagnetic field deep neural network model in step 4, and comparing the obtained predicted values ​​with the corresponding three components of the magnetic field in the test set. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the present invention;

[0029] Figure 2 It is the structural entity of a fully connected deep neural network;

[0030] Figure 3 Topological diagram of adding residual function to weights;

[0031] Figure 4 Schematic diagram of genetic algorithm optimization. DETAILED DESCRIPTION

[0032] First, those skilled in the art should understand that these implementations are only used to explain the technical principles of the embodiments of the present application, and are not intended to limit the protection scope of the embodiments of the present application. Those skilled in the art can make adjustments to them as needed to adapt to specific application scenarios.

[0033] In the description of the embodiments of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "connected" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments of the present application can be understood according to specific circumstances.

[0034] In the embodiments of the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0035] The present application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] See also Figure 1 to Figure 4 As shown, the embodiment of the present application discloses a geomagnetic modeling method based on a fully connected deep neural network, including:

[0037] Step 1. Obtain satellite magnetic data, and divide the satellite magnetic data into training set and test set after preprocessing; the characteristic elements of the satellite magnetic data include the time of reading the magnetic data, the longitude and latitude and the spherical center distance of the magnetic data, and the three components of the magnetic field in the NEC coordinate system; read the .cdf format file of the satellite magnetic data and convert the data into a .mat file; then filter the satellite magnetic data according to the interference label, local time condition and geomagnetic index condition of the satellite magnetic data; and normalize the filtered satellite magnetic data to solve the inconsistency of data units between the input layer and the output layer.

[0038] Step 2. Establish a geomagnetic field deep neural network model based on a residual network and a fully connected deep neural network; the geomagnetic field deep neural network model includes an input layer, four hidden layers and an output layer connected in sequence, the four hidden layers are connected in sequence, and the connection nodes between the input layer and the first hidden layer, between the first hidden layer and the second hidden layer, between the second hidden layer and the third hidden layer, between the third hidden layer and the fourth hidden layer, and between the fourth hidden layer and the output layer have weights, and a residual function and batch normalization processing are added to each of the weights; and the hyperparameters of the geomagnetic field deep neural network model are designed; wherein:

[0039] The input layer, hidden layer and output layer in the magnetic field deep neural network model respectively include multiple neurons, and each neuron in the input layer inputs a feature element ; The output of each neuron in the output layer The weights and activation functions are calculated by adding the residual function and batch normalization respectively; Figure 2 As shown, the activation function value of the input layer is , the activation function value of the first hidden layer is , the activation function value of the second hidden layer is , the activation function value of the third hidden layer is , the activation function value of the fourth hidden layer is , which is equivalent to the input feature vector; Indicates Layer The activation function value of a neuron, represents the activation function value of the first neuron in the first hidden layer, represents the activation function value of the first neuron in the second hidden layer, represents the activation function value of the first neuron in the third hidden layer, represents the activation function value of the first neuron in the fourth hidden layer, and so on. represents the fourth hidden layer The activation function value of a neuron, It is The activation function is calculated by the weighted sum of the activation function values ​​of all neurons in the layer; each output The corresponding neurons in the output layer The number of neurons in the hidden layer is calculated by the activation function. The range is first determined by an empirical formula, and then the relatively optimal number of neurons is finally determined by the cross-validation method. In this specific embodiment, the input layer is designed to have 3 neurons and the output layer to have 7 neurons.

[0040] Processing input data and output Add the residual function , and then constitute the identity mapping and skip connection of the geomagnetic field deep neural network model;

[0041] In this specific embodiment, adding the residual function specifically embeds the identity mapping directly into the network and transforms the learning objective into the residual function. , which means that the network no longer directly learns the target mapping, but learns the difference with the input, i.e., the residual , if a network layer is actually redundant, that is, no changes are required to the input data to achieve optimal performance, then the residual function of this layer is In theory, it will be automatically learned by the network to be 0; thus, the output of this layer becomes , an identity mapping is achieved.

[0042] Skip connections and residual learning strategies play a big role; residual learning (Residual Learning) Represents the data input to the network layer, Represents the output result, and since the neural network model is basically unable to adaptively learn the identity mapping function , embed the identity mapping directly into the neural network, so the problem becomes learning the residual function , the specific operation formula is as follows:

[0043] ;

[0044] Among them, if the network layer is determined to be a redundant layer, It will be automatically learned by the network, so that the identity mapping is constructed, and the performance of the network will not degrade. When residual learning is operated, the parameter values ​​in all network layers are greatly reduced, so that new features can be easily learned, which has a regularization effect on the network. Many residual units or residual blocks are stacked together to form a residual network, in which the residual units in the residual network not only have the same structure but are also continuous. The jump connection maps the output features of the upper layer to the end of the residual unit, and then adds them to the features of the residual output of the current layer. Finally, the sum is sent to the nonlinear activation function, and the result output from the activation function is regarded as the entire output result of the residual block. The jump connection is expressed by a specific formula below: ;

[0045] The identity mapping is equivalent to a short-circuit term. Figure 3 The networks shown are short-circuited one by one. represents the residual function of the first hidden layer, represents the residual function of the second hidden layer, represents the residual function of the third hidden layer, It represents the residual function of the fourth hidden layer. After short-circuiting, a new topological structure is formed. Multiple different network models are fused or connected in parallel, and the series structure of the original network at the bottom remains unchanged.

[0046] The output of the input layer or hidden layer Normalize to an array with a mean of 0 and a variance of 1. The specific method is: , , ,in is the dimension of the output, is the mean, is the standard deviation, It is 10 orders of magnitude smaller to prevent the occurrence of 0 which causes the division to fail. is the output of the i-th layer, is the normalized output.

[0047] The hyperparameters of the geomagnetic field deep neural network model include batch size, learning rate, reduction factor, training cycle Epochs and patience. This specific embodiment uses ReduceLROnPlateau to dynamically adjust the learning rate scheduler and make adjustments according to the performance changes of the model. When the monitoring index of a certain training cycle (epochs) (the MSE mean of all output data is used in this paper) stops decreasing, the learning rate will be multiplied by a reduction factor (Factor, a number less than 1) after the patience exceeds this waiting period, thereby reducing the learning rate. The update principle is shown in the following formula:

[0048] ;

[0049] Early stopping was used during model training, which is a technique commonly used to prevent overfitting of machine learning models, especially in geomagnetic field deep neural network models. For regression problems, after comparison, it was found that MSE directly measures the gap between the predicted value and the actual value, which is very consistent with the direct goal of the regression task - prediction accuracy. MSE has good mathematical properties, making it easy to handle in the optimization process. Adam optimizer is currently the best and most used optimizer. The selected activation function is ReLU function. The use of ReLU function helps the back propagation of gradients and facilitates the subsequent update of network parameters.

[0050] Step 3. Optimize the hyperparameters of the geomagnetic field deep neural network model based on mutation, crossover and selection of genetic algorithm; substitute the optimized hyperparameters into the geomagnetic field deep neural network model; specifically include:

[0051] Step 301. Establish the initial population corresponding to the number of samples (Batch Size), learning rate (Learning Rate), reduction factor (factor), training cycle Epochs and patience (patience), and set the parameter space of each population; the parameter space of this specific embodiment is set to Learning Rate [0.001, 0.01], factor [0.75, 0.95], patience [20, 30], Batch Size [32, 64], Epochs [150, 200];

[0052] Step 302. Each parameter value is regarded as a gene and represented in the form of binary coding; the parameter values ​​of the five hyperparameters, namely, the number of samples, the learning rate, the reduction factor, the training cycle and the patience, are combined to form a chromosome, and each chromosome represents an individual;

[0053] Step 303. Use the fitness function to calculate the fitness value of the individual. The fitness function is expressed by the mean square error between the predicted value and the true value of the fully connected network. The smaller the mean square error, the closer it is to the true value.

[0054] Step 304: Sort the fitness values ​​in the initial population, and screen out individuals whose fitness values ​​are less than a preset threshold, and use a genetic algorithm to perform crossover, mutation, and linear recombination to obtain the next generation population;

[0055] Step 305. For the next generation population, adopt steps 303 and 304, loop to obtain the T-th generation population, select the value with the largest fitness from the T-th generation population as the optimal parameter of each hyperparameter; and substitute it into the geomagnetic field deep neural network model.

[0056] Step 4. Input the training set into the geomagnetic field deep neural network model in step 3 for training; specifically including:

[0057] The training set of step 1 is input into the geomagnetic field deep neural network model of step 305 for training, and the loss is calculated and updated; wherein the loss adopts the mean square error between the predicted value and the true value of the previous propagation; the update includes a gradient descent process and a back propagation algorithm, and the Adam algorithm is adopted as the gradient descent algorithm, and the back propagation algorithm is adopted to update the weights of the connection nodes between the input layer and the hidden layer, between the hidden layer and the hidden layer, and between the hidden layer and the output layer in the geomagnetic field deep neural network model.

[0058] Step 5. Input the test set into the trained geomagnetic field deep neural network model for verification, specifically including: input the test set into the trained geomagnetic field deep neural network model in step 4, and compare the obtained predicted values ​​with the corresponding three components of the magnetic field in the test set.

[0059] The present invention compares the root mean square error of the fully connected network, the residual network optimized fully connected network, and the residual network combined with the genetic algorithm to optimize the fully connected network in the NEC coordinates in the three directions of X, Y, and Z, as shown in Table 1:

[0060] Table 1 Root mean square error comparison analysis table

[0061]

[0062] As shown in Table 1, the root mean square error values ​​of the fully connected network jointly optimized by the residual network and the genetic algorithm in the three directions of X, Y, and Z are all very small, thereby ensuring the accuracy of the optimized model in the learning task while avoiding the problem of gradient vanishing or gradient exploding in the model.

[0063] In the description of the embodiments of the present application, it should be noted that in the description of the present application, terms such as "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is only for the convenience of description, and does not indicate or imply that the device or component must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present application.

[0064] In the description of the present application, the description with reference to the terms "one embodiment", "some embodiments", "in the present embodiment", "specific example", or "some examples" etc. means that the specific features, mechanisms, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, mechanisms, materials or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0065] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A geomagnetic modeling method based on a fully connected deep neural network, characterized in that: include: Step 1. Obtain satellite magnetic survey data, and divide the satellite magnetic survey data into training set and test set after preprocessing; Step 2. Establish a geomagnetic field deep neural network model based on a residual network and a fully connected deep neural network; the geomagnetic field deep neural network model includes an input layer, four hidden layers and an output layer connected in sequence, the four hidden layers are connected in sequence, and the connection nodes between the input layer and the first hidden layer, between the first hidden layer and the second hidden layer, between the second hidden layer and the third hidden layer, between the third hidden layer and the fourth hidden layer, and between the fourth hidden layer and the output layer have weights, and a residual function and batch normalization processing are added to each of the weights; and the hyperparameters of the geomagnetic field deep neural network model are designed; Step 3. Optimize the hyperparameters of the geomagnetic field deep neural network model based on mutation, crossover and selection of genetic algorithm; Substitute the optimized hyperparameters into the geomagnetic field deep neural network model; include: Step 301. Set the parameter ranges of the number of samples, learning rate, reduction factor, training cycle and patience, randomly select values ​​of the five hyperparameters to form individuals, and randomly select multiple individuals to form the initial population; Step 302. Each parameter value is regarded as a gene and represented in the form of binary coding; the parameter values ​​of the five hyperparameters, namely, the number of samples, the learning rate, the reduction factor, the training cycle and the patience, are combined to form a chromosome, and each chromosome represents an individual; Step 303. Use the fitness function to calculate the fitness value of the individual. The fitness function is expressed by the mean square error between the predicted value and the true value of the fully connected network. The smaller the mean square error, the closer it is to the true value. Step 304: Sort the fitness values ​​in the initial population, and screen out individuals whose fitness values ​​are less than a preset threshold, and use a genetic algorithm to perform crossover, mutation, and linear recombination to obtain the next generation population; Step 305. For the next generation population, adopt steps 303 and 304 to loop to obtain the T generation population, select the value with the largest fitness from the T generation population as the optimal parameter of each hyperparameter; and substitute it into the geomagnetic field deep neural network model; Step 4. Input the training set into the geomagnetic field deep neural network model in step 3 for training; Step 5. Input the test set into the trained geomagnetic field deep neural network model for verification.

2. The geomagnetic modeling method based on a fully connected deep neural network according to claim 1, characterized in that: The characteristic elements of the satellite magnetic data in step 1 include the time of reading the magnetic data, the longitude and latitude and the distance from the center of the sphere of the magnetic data, and the three components of the magnetic field in the NEC coordinate system; the pre-processing of the satellite magnetic data in step 1 is specifically to filter the satellite magnetic data according to the interference label of the satellite magnetic data, the local time condition, and the geomagnetic index condition; and the filtered satellite magnetic data is normalized.

3. The geomagnetic modeling method based on a fully connected deep neural network according to claim 1, characterized in that: The input layer, hidden layer and output layer of the geomagnetic field deep neural network model in step 2 respectively include multiple neurons, and activation functions are respectively provided in the input layer, hidden layer and output layer. Each neuron in the input layer respectively inputs feature elements After adding the residual function and batch normalization processing to calculate the weights and nonlinear activation functions, the output of each neuron in the output layer is .

4. The geomagnetic modeling method based on a fully connected deep neural network according to claim 3 is characterized in that: In step 2, each of the weights is added with a residual function and subjected to batch normalization processing, specifically including: The output of the input layer and each hidden layer is , when processing input feature elements And each out Add the residual function , expressed as ; Construct the identity mapping of the geomagnetic field deep neural network model; Each output Normalize to an array with a mean of 0 and a variance of 1. The specific method is: , , ,in is the dimension of the output, is the mean, is the standard deviation, It is 10 orders of magnitude smaller to prevent the occurrence of 0 which causes the division to fail. is the output of the i-th layer, is the normalized output.

5. The geomagnetic modeling method based on a fully connected deep neural network according to claim 1, characterized in that: The hyperparameters of the geomagnetic field deep neural network model designed in step 2 include the number of samples, learning rate, reduction factor, training cycle and patience.

6. The geomagnetic modeling method based on a fully connected deep neural network according to claim 1, characterized in that: The step 4 specifically includes: inputting the training set into the geomagnetic field deep neural network model in step 3 for training, specifically includes: inputting the training set in step 1 into the geomagnetic field deep neural network model in step 305 for training, calculating the loss and updating; wherein the loss adopts the mean square error between the predicted value and the true value of the previous propagation; the updating includes a gradient descent process and a back propagation algorithm, adopting the Adam algorithm as the gradient descent algorithm, and adopting the back propagation algorithm to update the weights of the connection nodes between the input layer and the hidden layer, between the hidden layer and the hidden layer, and between the hidden layer and the output layer in the geomagnetic field deep neural network model.

7. The geomagnetic modeling method based on a fully connected deep neural network according to claim 1, characterized in that: The step 5 of inputting the test set into the trained geomagnetic field deep neural network model for verification specifically includes: inputting the test set into the geomagnetic field deep neural network model trained in step 4, and comparing the obtained prediction value with the corresponding three components of the magnetic field in the test set.